A Bayesian Framework To Study Individual Differences In Semantic Representations

Researchers have now started using Bayesian hierarchical modeling to study semantic representations. Ramotowska et al. (2024) have contributed to the further development of this methodology by examining natural language quantifier expressions such as most, few, etc. Their model disentangles three key semantic parameters: the meaning threshold of quantifiers, the uncertainty surrounding that threshold, and response noise. In this work we pursued two main goals. First, we refined the original model by (1) incorporating informative prior distributions that generate meaningful predictions, (2) adding a covariance structure to broaden its scope to study semantic representations across time and paradigms, and (3) changing its link function to clearly identify quantifiers' truth conditions. Second, using our model, we aimed to experimentally test the stability of semantic representations across time and paradigms, and to evaluate two competing theoretical approaches that explain quantifier vagueness. For that, we applied our model to two datasets. We reanalyzed data (n=63) from Ramotowska et al. (2023) to test the stability across time. We then conducted a new experiment (n=178) to test the stability across paradigms. Our results suggest that semantic representations change over time and paradigms. Moreover, we find that uncertainty around the meaning threshold is observable at both the group and individual levels, which supports the threshold model of meaning over alternatives based rigidly on logic and Generalized Quantifier Theory. Our proposed model allows researchers to study semantic representation of quantifier meanings effectively and offers the versatility to be applied to a wide range of linguistic constructs.